US 20080140597 A1
The linehaul optimization algorithm is a hybrid algorithm incorporating the basic structure of the Hungarian Assignment algorithm or equivalent and an improvement heuristic. However, in this particular case there is not a direct application of the existing, well-known Hungarian Assignment algorithm or equivalent in conjunction with a heuristic. The structure of the assignment model being used here is based on the Hungarian Assignment algorithm or equivalent but has been revised and enhanced to identify systematically (by using grouping) a number of equivalent optimal solutions (that give the same optimal, minimum cost for a particular iteration). A heuristic is then used to identify the ‘best’ optimal solution, of the many identified, that would contribute the most incremental cost reduction in future iterations of the heuristic. There is no method at present that finds all possible combinations using the Hungarian Assignment algorithm or equivalent because for most applications any optimal solution is sufficient. However, this optimization model uses an iterative procedure where the outcome of each iteration affects the next iteration significantly and a simple application of the original Hungarian Assignment algorithm or equivalent is not acceptable and does not provide the overall optimum. At each iteration, a leg assignment is made (where a trailer is moved over more than one leg) or a reassignment is made to move a trailer in a different time period by either an owner-operator (i.e. company) tractor or a broker (i.e. outside company) tractor. Furthermore, the original Hungarian Assignment algorithm or equivalent is not designed to solve a scheduling problem; it is designed to solve an assignment problem. In other words the algorithm has been revised and combined with a heuristic to solve a scheduling problem.
1. A method suitable for optimizing a linehaul operation, the method comprising the steps of:
(a) providing at least one assignment algorithm designed to solve an assignment problem;
(b) providing at least one heuristic;
(c) combining the at least one heuristic with the at least one assignment algorithm to solve the scheduling problem, the combining step comprising the following acts:
(i) identifying at least one equivalent assignment solution to an instantaneous assignment problem using the at least one assignment algorithm; and
(ii) using the at least one heuristic to select one assignment solution from the at least one equivalent assignment solution such that the selected one assignment solution reduces the cost of at least one other instantaneous assignment solution.
2. The method as recited in
3. The method as recited in
4. The method as recited in
5. The method as recited in
6. The method as recited in
7. The method as recited in
8. The method as recited in
9. The method as recited in
10. The method as recited in
11. The method as recited in
12. The method as recited in
13. The method as recited in
14. The method as recited in
15. The method as recited in
16. The method as recited in
17. The method as recited in
18. The method as recited in
19. The method as recited in
20. The method as recited in
21. The method as recited in
22. The method as recited in
23. The method as recited in
This application claims priority from Canadian Patent Application Number 2,551,467 filed on July 4, 2006.
This application relates to linehaul operations in general, and to a system and method suitable for optimizing linehaul operations, in particular.
The efficiency of transportation operations is one of the factors having a significant effect on the country's economic growth. A linehaul operation, one of the biggest contributors to the total operating cost of a trucking company, involves moving full loads of less than unit load or full unit load shipments using owner-operator and/or company units.
A review of the prior art revealed that for an optimization procedure to be successful, the following steps have to be executed in sequential order:
The problem must be defined in the most appropriate manner such that it meets the practical requirements of the real-world problem. This ensures that, when the model is developed, it is a more accurate mathematical representation of the problem. The algorithm developed to meet the requirements of the model can be an existing algorithm, a modified existing algorithm, or a new algorithm developed specifically to solve the model. Since algorithms are usually not developed to solve a specific problem, they require significant modification to address specific operational requirements. If the problem to be solved is a unique problem and cannot be addressed by any of the existing algorithms, it is necessary to develop a new algorithm to meet its requirements.
Some of the existing modeling and solution approaches from recent research that are relevant to linehaul operations are discussed in the following sections of this chapter. There are also some commercially available software packages intended to assist in modeling transportation operation problems, and/or finding a solution to the modeled problem. The advantages and disadvantages of these packages are also discussed in the Commercial Modeling Packages section of this chapter. Although considerable research has been completed in the transport operations area, the scale, nature, and complexity of the current linehaul problem addressed in the present invention is significantly different from ones that have been studied so far and reported in prior art.
The linehaul planning problem and the Vehicle Routing Problem (VRP) in general are non-deterministic polynomial time hard (NP-hard) combinatorial optimization problems for which no efficient algorithms are known that provide a global optimum solution. The solution approaches to NP-hard combinatorial problems can be grouped into two categories, namely exact solution approaches and heuristic solution approaches. The exact solution approach looks into almost every possibility to find the optimum solution and is useful for problems of relatively small size. Heuristic algorithms do not indicate how good the result is in relation to the global optimum, but they are the most practical choice for solving large problems because of the efficiency gains and improvements that they can potentially make.
The Traveling Salesman Problem (TSP) is one of the best known problems in combinatorial optimization (Cook et al., 1998). TSP is a sub-problem of the VRP (Lawler et al., 1992). VRP can have a larger set of solution possibilities, depending on the nature of the problem. The TSP can be described as having N destinations, and the cost, Cij to travel from destination i to j. A salesman (vehicle) starts from the depot to visit each destination once and return back to the depot. The problem is to find the optimal route and the destination visiting order that would yield the minimum total travel cost. If the entire set of solutions were enumerated, the number of calculations would be N! (Taniguchi et al., 2001). Table 2 shows the rapid growth of computational time for solving TSP when N, the number of destinations (customers), increases. In this example, a 10 gigahertz (GHz) processor with basic computation time of 10.10 seconds was used. In an enumeration method, one step may take several computations, but to be conservative, it was treated as if it would take the same time as for one computation. As the following table shows, when N increases from 10 to 50 destinations, computation time to find an exact solution increases from 0.00036 seconds to 9.64*1037 billion years:
Clarke and Wright (1964) developed the earliest heuristic algorithm (that is the Clarke-Wright heuristic algorithm) to solve the VRP. Significant research done to solve the VRP has helped in the development of many heuristic algorithms, which are now commonly used, such as Tabu Search (TS), Simulated Annealing (SA), Genetic Algorithms (GA), and Repeated Descent (RD). Gendreau et al. (1997) applied several heuristics to the VRP and their findings were that the GA, SA, and TS techniques performed significantly well in VRP compared to other algorithms. Wright and Marett (1996) compared the performance of RD, SA, and TS on the TSP and TS outperformed both SA and RD.
Taniguchi et al. (1998) compared the performance of the GA, SA, and TS when applied to the VRP, and concluded that the TS reached the best known solution with the shortest computation time. Reviews on ten of the most common TS heuristics for the VRP with their main features of neighbourhood structures, short term memory, long term memory, and intensification can be found in one of the most recent surveys conducted by Cordeau and Laporte (2002).
The classical VRP is to determine the routes and service order of customers with known demands from a central depot. Although classical VRP models and algorithms cannot be applied to solve linehaul scheduling problems, due to some similarities in the nature of their operations, a survey of the literature pertaining to classical VRP is also included in this chapter. This was done for the purpose of ascertaining what scholars have done to handle similar situations in this subject. It also revealed an understanding of current active research modeling approaches and solution algorithms.
A comprehensive literature survey on the classical VRP and some of its variations was made by Fisher (1995), where he categorized VRP methods into three generations. The first generation consisted of simple heuristics developed in the 60's and 70's, which were mainly based on local search or sweep methods. The second generation included mathematical programming based heuristics (near-optimization algorithms), which differ from normal heuristics in that they include the generalized assignment problems and set partitioning to approximate the VRP. The third and the last generation is the one currently being studied intensively and includes exact optimization algorithms and artificial intelligence methods such as TS, GA, and SA. Christie and Satir (2004) reviewed several existing case studies and found that a 10% to 30% reduction in operational costs is possible through VRP optimization for P & D operations. Satir (2003) analyzed the P & D operations (at a terminal) of a major trucking company in Atlantic Canada, and compared the potential efficiency gains of implementing an optimization system. The results of the case study showed that operational costs could be reduced by as much as 50%.
The classical VRP is not applicable (without additional constraints and considerable modification) to real-world transportation operation problems. One of the problems closest in nature to the classical VRP is that of parcel P & D. In most P & D operations vehicles have various fixed capacities and can serve customers (with varying demands) only within this capacity. This type of problem is referred to as the Capacitated VRP (CVRP) in literature. In most real-world transportation problems, customers have certain time windows within which they have to be served by a fleet of vehicles of varying capacities. This type of CVRP with time windows (TW) is known as VRPTW. The VRPTW has many practical applications such as in LTL P & D from a central depot or multiple depots. Tan et al. (2000) investigated various third generation heuristic methods such as TS, SA, and GA to solve the VRPTW.
Although the classical VRP and some of its variations have been studied extensively in literature, VRP with FTL (VRPFTL) has received relatively less attention in literature. The closest VRP variations to VRPFTL are the multi-depot vehicle scheduling problem with time windows (MDVSPTW) and the multi-trip vehicle routing and scheduling problem (MTVRSP), which have been studied comparatively more. The main difference between VRPFTL and MDVSPTW is that while VRPFTL has one FTL demand per trip (one customer), MDVSPTW can have one or more customer demands to fill the truck per trip. Brandao (1997) points out that the MTVRSP has additional constraints to the VRPTW such as a vehicle can make more than one trip, vehicle's capacity is considered in terms of both volume and weight, the access to some customers is restricted to some vehicles, and driver's schedule must comply with the driving regulations. Desaulniers et al. (1998) published a paper on MDVSPTW and waiting costs, which bears some similarities to the problem addressed in this present invention. A literature survey was also made by Desrosiers et al. (1995) for multi-depot VRP related problems. A recent VRPFTL paper published by Arunapuram et al. (2003) introduced new techniques that were intended to solve such problems with exact solutions, which use a branch-and-bound algorithm to solve an integer-programming formulation of the VRP.
Yang et al. (2004) published a paper which considered various costs associated with empty miles, delayed completion times, and job rejection when using real time information to improve productivity. Liu et al. (2003) developed a heuristic algorithm to determine a mixed truck delivery system that allows both hub-and-spoke and direct shipment delivery modes. Smilowitz et al. (2003) studied the possible integration of longhaul operations for package delivery services, which has similar context to some of the components of the problem addressed in the present invention.
Trucking operations consist of P & D and linehaul operations. The classical VRP is applicable primarily to P & D operations. Although a considerable amount of research on the classical VRP exists in literature, relatively less research has been done on vehicle scheduling and planning for linehaul operations. Planning and scheduling of transportation operations has potential applications in the rail, maritime, trucking, and airline modes of transportation. Some of the real-world problems, models proposed, applications, and implementation benefits pertaining to planning and scheduling in literature are discussed in this section.
Linehaul planning is more challenging than the typical VRP in that the fleet of vehicles has to satisfy each terminal's varying demands from every other terminal, whereas in the classical VRP, it has to satisfy customer demands only from a central terminal. One of the most similar problems to linehaul planning is that of airline planning and scheduling. Significant research has been undertaken in this area, and OR has had a significant impact on the airline industry. Clarke and Smith (2004) point out that American Airlines had an estimated cost savings of approximately $18 million USD per year (relative to previously used enumeration methods) by implementing an optimization-based crew pairing system in 1989. Cook (2000) claims that American Airlines testified that over $500 million was generated annually in incremental profits due to their optimized schedule planning system. In a similar manner, when the Delta Airlines first implemented their fleet assignment module in 1992, the planning group reported an approximate cost savings of more than $100,000 USD per day (Subramaniam, 1994). These cost savings increased as planners gained more experience and confidence with the module, and reached an estimated $200,000 USD per day in late 1993. Ryan et al (2001) reported not only an increase in savings of more than $15,655,000 NZ per year in crew scheduling costs at Air New Zealand, but also an increase in crew member satisfaction.
Many of the problems found in maritime, railway, and airline transportation operations are based on multi-commodity flow problems. A multi-commodity flow problem is defined in the Algorithms and Theory of Computation Handbook as “a maximum-flow problem (finding the maximum flow between any two vertices of a directed graph) involving multiple commodities, in which each commodity has an associated demand and source-sink pairs”.
Barnhart et al. (1995) published a partitioning solution procedure to address large-scale multi-commodity flow problems with many commodities. Barnhart et al. (2000) presented a model and iterative solution approach to solve the problem of determining the type of aircraft to assign to each flight and the exact departure time of each flight, given the set of flights with their time windows for a large U.S airline. It should be noted that the process of fleet assignment in airline operations is very similar to tractor assignments in trucking operations, subject to different operational and regulatory constraints.
Barnhart, Hane, and Vance (2000) proposed a branching rule to find a heuristic solution and compared branch-and-price and branch-and-price-and-cut methods to find optimal solutions for highly capacitated problems, specific to telecommunication applications. Barnhart et al. (2002) proposed a new formulation and solution approach that captures network effects and involves the profit maximizing assignment of aircraft types to flight legs.
The general framework for aircraft and crew schedule and planning is given in (Barnhart, 2004). Although decomposing the problem into four subproblems and optimizing them individually would yield a less efficient solution than an overall optimization approach, it is necessary to decompose problems of large size such as linehaul problems in a similar manner to obtain a practical solution in a feasible time.
Toth (2004) stresses the need to undertake planning activities in an efficient way for railway transportation operations and the importance of using computer-aided tools to improve the planning ability of railway systems. He points out some of the problems arising in railway optimization to be line planning, platforming, shunting, locomotive assignment and scheduling, crew scheduling, and train timetabling. In particular, the train timetabling problem, which deals with train scheduling and dispatching, is of similar nature to linehaul planning. Caprara et al. (2002) proposed a graph theoretic formulation to model and solve the train timetabling problem using a directed multigraph in which nodes correspond to departures or arrivals at a certain station at a given instant.
It is also beneficial to understanding the recent research in maritime transportation, since ship routing and scheduling problems are related to linehaul planning problems. Christiansen et al. (2004) reviewed the current status of ship routing and scheduling based on literature published in the last decade. They presented perspectives regarding future developments and use of optimization-based decision-support systems for ship routing and scheduling. They also examined tactical and operational fleet planning operations and consider problems that comprise various ship routing and scheduling aspects. Hooghiemstra et al. (1999) proposed a model for a real ship planning problem, which is a combined inventory management problem and a routing problem with time windows. They also discuss model adjustments to the proposed model to decompose the problem into ship routine and inventory management subproblems. Flatberg et al. (2000) combined an iterative improvement heuristic with an exact solution approach to minimize transportation costs (while ensuring satisfactory inventory levels) within acceptable time limits. The model was for a real-world problem of transporting a single commodity between producing and consuming factories within the same company using a fleet of vessels.
Powel (1996) proposed a stochastic formulation of the dynamic assignment problem with an application to truckload motor carriers where demands arise randomly and continuously throughout the day. Powel et al. (2002) did a case study on implementing operational planning (real-time load matching) model for a trucking carrier. Powel et al. (2000) found that the value of global optimization is lost when users ignore the solution suggested by the model. Crainic and Laporte (1997) discussed some of the main issues in freight transportation operations planning and presented appropriate Operations Research models and methods, as well as computer-based planning tools. They examined the strategic, tactical, and operational decision-making levels and also reviewed significant methodological and instrumental developments in these areas.
The earliest models in freight transportation were deterministic models that captured the time staging of physical activities as against the time staging of information (Powell, 2003). Powell and Carvalho (1998) proposed the use of linear functional approximations to measure the future impact of current decisions in freight transportation planning. Godfrey and Powell (2002) found that non-linear functional approximations produced a better and more stable solution than linear approximations.
Powell et al. (2004) explained a transportation system in terms of the management of the static resources (such as terminals) and passive resources (such as people and equipment) that constrain the system. They also introduced a basic resource model for modeling resources and their attributes. Powell, Shapiro, and Simaõ (2002) formulated a large-scale driver scheduling problem as a multistage dynamic resource allocation problem for an LTL trucking application. The large scale of the problem rendered the use of commercial solvers infeasible. They demonstrated that their techniques provided high quality solutions within reasonable time limits. Powell and Topaloglu (2005) discuss the fleet management problem, which involves managing fleets of equipment to meet the customer requests that arrive randomly over time, often requiring service within a small time window.
Transportation operation problems can be grouped into point-to-point (node) routing and arc routing problems. While in node routing problems the service activity occurs at all (or at some subset of) the nodes, in arc routing problems a single vehicle or a fleet of vehicles must service all (or some subset of) the arcs (and/or edges) of a transportation network (Corberan et al. 2005). The majority of operations fall under the category of point-to-point routing, such as VRP and linehaul operations. Arc routing applications include garbage collection, snow removal, sweeping, gritting, mail delivery, meter reading, school bus routing etc. (Eiselt et al. 1995). Although there are many commercially available packages for arc routing, these packages cannot be used on the current problem, because it is a point-to-point routing problem.
TransCAD is the only software package that fully integrates GIS with demand modeling and logistics functionality (TransCAD 2006). It can be used to solve a variety of logistics problems such as vehicle routing/dispatching, arc routing, and network flow and distribution analysis. Satir (2003) did a case study for real-world P&D operations, using TransCAD to model and solve the problem. The results showed that TransCAD is a very effective scheduling tool for the VRP. However linehaul planning and scheduling problems cannot be solved by using this package.
The ILOG optimization packages are the most popular and considered the de facto optimization standard in OR. The main software components of these optimization packages are: ILOG CPLEX, ILOG Solver, ILOG Dispatcher, and ILOG Scheduler. ILOG (2003) discusses the various components in detail in their respective manuals. A brief description of each of these components from these manuals is summarized in this section.
ILOG CPLEX consists of C, C++, Java, and C# libraries that solve Linear Programming (LP) problems and its extensions such as network flow problems, quadratic programming (QP) problems, quadratically constrained problems, and Mixed Integer Programming (MIP) problems. In MIP problems, any or all of the LP or QP variables are further restricted to take integer values in the optimal solution.
ILOG Solver is a C++ library that allows the user to model optimization problems independently of the algorithms used to solve the problem. The two main techniques for solving the optimization problems are search strategies (Tabu Search or guided local search) and constraint propagation. A comparative study of eight constraint programming languages done by Fernandez and Hill (2003) showed the ILOG Solver to be the fastest, most efficient, and robust.
ILOG Dispatcher is a C++ library that offers features especially designed to solve problems in vehicle routing and maintenance-technician dispatching, This library is made up of classes that represent various aspects of routing plans, their vehicles, visits, and constraints. The Dispatcher finds various applications such as modeling a vehicle routing problem, solving a vehicle routing problem, modeling P & D operations, and routing multiple tours per vehicle. A solution to a routing problem involves the three stages of describing the problem, modeling the problem, and lastly solving the problem by using local search The Dispatcher can also be used to solve a VRP such as an assignment problem, or a P & D problem.
ILOG Scheduler is a C++ library intended to primarily help in solving scheduling and resource allocation problems. It works in conjunction with ILOG Solver and provides specialized modeling and algorithmic enhancements for scheduling problems. ILOG Scheduler makes it feasible to also take temporal and capacity constraints into consideration. The two types of temporal constraints that can be added are precedence constraints and time-bound constraints. While precedence constraints are used when the user needs to specify when an activity must start or end with respect to the start or end time of another activity, time-bound constraints are used to specify when an activity must start or end with respect to a given time.
ILOG CPLEX is not a practical choice for the problem at hand, because the problem requires a MIP formulation, which is extremely processor-intensive and time consuming for a problem of this size. ILOG Dispatcher could not be used for this problem either, because the Dispatcher is specifically designed for P & D type operations.
Using the ILOG Scheduler in conjunction with the ILOG Solver would have been the most appropriate choice to model and solve the problem. However, the problem at hand is not a pure scheduling problem, and has regulatory constraints and operational preferences in addition to the regular operational constraints. The pre-designed modeling environment would make the task of taking these additional regulatory constraints and operational preferences into account very cumbersome.
The latest product from ILOG, which is probably the most appropriate optimization package for linehaul planning is called Transport PowerOps, and was released as recently as April 2005. This package helps planners determine the lowest cost solutions for assigning vehicles, loading shipments and sequencing routes (ILOG PowerOps, 2006). The most important features of PowerOps are that it can handle hours of service regulations, driver rest rules, and multiple time windows. However, a shortcoming of this package is that the current problem requires different fleets to be subjected to different service regulations and for team and single driver vehicles to be handled differently. Some of the other disadvantages of using a package such as PowerOps (besides the company's potential dependency on an external package) are that the model and the operational requirements will have to be updated (building a new model) constantly to accommodate any changes or enhancements made to newer versions of the package, and this would make it difficult to integrate it with the existing IS. Operational preferences which cannot be taken into account by this package at present can be embedded using heuristics. However, this would require significant amount of work in addition to modeling the problem, which by itself is a tedious task. Although this package lacks some features and is not an off-the-shelf package (thus requiring specialized knowledge and expertise to make use of it), it is still a valuable asset to model and optimize linehaul operations.
There is a lack of software specifically designed for linehaul planning and scheduling.
Planning for large scale linehaul operations is one of the most challenging transportation operations problems due to the complexity (number of decision variables involved) and nature of the operations. The dynamic nature of linehaul operations makes it stressful for planners to make operational decisions and manual planning is a cumbersome and error-prone task.
The present invention addresses the problem of linehaul planning by developing a decision support system that is aimed primarily at assisting planners in making the daily operational decisions, and also to help management in making tactical and strategic decisions. One of the main advantages of this integrated system is that it is independent of any commercially available optimization packages and is customized specifically for linehaul operations. In most of the research done on optimization, dependence on other optimization software or programs is a hindering factor, as customization is a very cumbersome process.
One objective of this invention was to develop a customized solution procedure for real-world transportation problems. In order to ensure that the solution developed could be used in practical situations, a company case study was pursued. This case study made it possible to obtain the relevant data and inside information on linehaul operations and planning, which made it feasible to quantify the potential operational efficiency gains.
Although embodiments of the decision support system relate specifically to a company's linehaul operations, the framework disclosed is such that it would be obvious for a person of skill in the appropriate art to easily adapt and modify what is disclosed herein to suit any linehaul operation.
The efficiency gains are quantified in terms of reduction in linehaul operating costs from the existing operating costs. The linehaul operating costs have four major contributors and are obtained by combining the full mile cost, empty mile cost, broker mile cost, and wait cost. One of the non-quantifiable efficiency gains is higher customer satisfaction, which results from providing them with better service. The other non-quantifiable benefits include higher planner satisfaction because most of the operational decisions are made by the optimization system and higher owner-operator satisfaction, which is achieved by giving them jobs that best suit their work expectations.
Some teachings made by the present disclosure are, inter alia, as follows:
There are no commercially available software packages that can duplicate complex linehaul operations. There was a real need among transportation companies to ease the burden of cumbersome manual planning by using a computer program that would assist in making the linehaul operational decisions. Since the trade-offs are made automatically by the program, it eliminates human errors, which leads to significant efficiency savings.
The Integrated Optimization System developed in accordance with one embodiment of this invention contains all the required modules in one single computer program. It is typical to be dependent on different software package(s) to address complex transportation optimization problems. The uniqueness of this optimization system is that it is independent of any other software packages and offers a very comprehensive solution to linehaul operational issues.
In order to automate the process of making the trade-off between hiring broker tractors and owner-operator tractors, it was necessary to add dummy jobs as well as broker tractors to the matrix used in the Hungarian Assignment algorithm or equivalent. In a typical Hungarian Assignment or equivalent procedure, dummy cells are added either to the rows or columns, but not to both.
The size of the matrix used in the Hungarian Assignment algorithm or equivalent is also reduced in every iteration to make the iterative procedure as efficient as possible. The jobs that are eliminated in the current iteration are jobs for which the broker tractor cost is lower than the owner-operator tractor cost and those jobs that have been completed.
The development of the iterative procedure is a major contribution and has introduced many unique concepts such as that of relative time, time step, and also an algorithm to apply the iterative procedure efficiently for the current problem.
There is no known algorithm in literature to determine all the possible optimum solutions for the Hungarian assignment or equivalent problem. The potential efficiency gains encouraged the development of an algorithm to determine and group all possible optimum solution combinations, where every optimum solution combination gives the equivalent lowest total operating cost.
In typical assignment problems, there is no need to find all the possible optimum solutions as any optimum will suffice. However, it was necessary to introduce the best optimum concept because of the complexity of the solution procedure developed and used in this invention. Although any optimum solution would be sufficient for use in a non-iterative procedure, the algorithm developed to determine best optimum solution is a very important contribution because of the cumulative effect it would have on the total operating cost in an iterative procedure.
It would be impossible to manually optimize the values of the weighting coefficients used in the Optimization Model within a feasible length of time and with a reasonable amount of effort. The development of an automated modification procedure to perform this optimization is an important contribution, because it also makes it possible to run this module on a regular basis when further modification of the coefficients is required such as with an operational policy change or the addition of terminals or switch points.
According to one aspect of the present invention, there is provided: a method suitable for optimizing a linehaul operation, the method comprising the steps of: (a) providing at least one assignment algorithm designed to solve an assignment problem; (b) providing at least one heuristic; (c) combining the at least one heuristic with the at least one assignment algorithm to solve the scheduling problem, the combining step comprising the following acts: (i) identifying at least one equivalent assignment solution to an instantaneous assignment problem using the at least one assignment algorithm; and (ii) using the at least one heuristic to select one assignment solution from the at least one equivalent assignment solution such that the selected one assignment solution reduces the cost of at least one other instantaneous assignment solution.
Other aspects and features of the present invention will become apparent to those ordinarily skilled in the art upon review of the following description of specific embodiments of a system and method suitable for optimizing linehaul operations in conjunction with the accompanying drawing figures.
Embodiments of the present invention will now be described, by way of example only, with reference to the accompanying drawing figures, wherein:
Like reference numerals are used in different figures to denote similar elements.
As used in the present description, appended claims, and drawings, the term unit refers to a unit of transport such as for example a container, trailer, or car load. The use of the term trailer or tractor or equivalent in the examples is illustrative only for applicability to the trucking mode of transportation and in no way limits the applicability of the present invention to other modes of transportation than trucking. The invention finds applicability to linehaul operations such as but not limited to trucking, busing, navigation, aviation, rail, and any other mode of transportation and combinations thereof. Furthermore, the use of the two countries Canada and United States in the specification is exemplary only. Application of the invention to other countries is obvious in view of the present specification, drawings, and claims.
As used in the present specification, appended claims, and drawings, the following abbreviations respectively mean:
The mathematical definition of the problem of linehaul planning and scheduling is presented as follows.
There is an undirected traffic network G=(V, E), with the following parameters:
A feasible solution consists of:
Subject to the following operational constraints:
The objective is to find a feasible solution that minimizes the total operating cost, where the operating cost is a function of owner-operator's full miles, brokered full miles, wait time of owner-operator tractors for jobs at non home-base nodes.
In order to make the most of the Optimization Model, the following sequence of steps was executed: Mathematical definition of the problem, Model Development, and Solutions Algorithms. This customized solution development for linehaul operations enabled the Optimization Model to provide more efficient operational decisions.
The Optimization Model focuses on producing a better solution, which yields lower operating costs than known models. The Optimization Model also makes use of the same information as that available to dispatchers and planners, but can handle very complex calculations. It is an iterative Hybrid Model, where exact solution assignment algorithms are used in conjunction with improvement heuristics in every iteration to improve the final solution.
The Optimization Model uses the following structures which are known from other models:
The following additional structure (leg_arr) introduced to handle switch points. This structure stores the leg information for any pair of terminals
Referring to the drawings,
The systematic approach followed to make the job assignments to tractors using this model is given in the steps and acts recited below:
Every job in the Optimization Model is thus treated as having a fixed number of legs, with one leg processed per iteration.
The broker cost for each leg of the load is calculated by using the current broker cost per mile, and the distance between the origin and destination of that leg, and is stored in the broker cost matrix. Each cell of the cost matrix represents the cost that a trucking company would incur if their tractor was to move a particular load (either for the leg currently being examined, or for the move from origin to destination in the case that there are no switch points between the origin and destination of that load). This cost is calculated using certain formulae, depending on several criteria which are discussed in greater detail further below. When it is not viable or beneficial to move a company tractor to haul a load, the corresponding cells are assigned a big number indicating that such a move is not feasible. These cases are as follows:
If the cost of assigning a job to an owner-operator tractor (from the cost matrix) is lower than the cost of assigning that job to a broker tractor (from the broker cost matrix), this job is considered assigned to a broker tractor, and unavailable for that iteration. This is done in order to reduce the size of the matrix sent to the Hungarian Assignment algorithm or equivalent, and save processing time.
The Hungarian Assignment algorithm or equivalent is one of the most efficient exact solution assignment algorithms. It can be used on square matrices to find the optimum assignments, which yield the overall minimum operating cost. The assignment of a job to a tractor is indicated by a “1”, and no assignment is indicated by a “0” at the corresponding cell of the decision matrix, X. The cost matrix is transformed to a matrix known as the effective matrix, in the Hungarian Assignment algorithm or equivalent. This matrix is used in the find optimum combinations function, where all possible optimum combinations are grouped in such a way that changing a particular job's assignment to a certain tractor within that group still yields the same optimum solution. The different combinations found are then analyzed in the final optimal assignments function to determine the best optimum combination of job assignments to tractors. The difference between the best optimum combination and any possible combination is that, although the operating cost for the current iteration is the same for both, the best optimum combination yields lower operating cost for the next iteration. Using best optimum combinations in every iteration accumulates significant savings at the end of the iterations, thus yielding a solution that is closer to the global optimum solution. Possible optimum combinations are weighted to determine the best optimum solutions with factors such as wait time cost, home base and getting closer to home base factors (the tractors should be at the right location in order to lower the costs for the next iteration), and fairness factors such as a tractor's maximum time away and maximum mileage.
If at least one assignment is made to a company tractor in the current iteration, another iteration is performed with the remaining unassigned loads (including loads that were assigned to brokers in the current iteration). This process is repeated until no more company tractor assignments can be made. In the case where all possible assignments have been made, the program terminates by printing all the required output.
The coefficients of the weighting factors used in the final optimal assignments function had to be determined to yield the planning solution with the lowest possible operating cost. It was a very cumbersome and almost impossible task to determine the optimum values for these coefficients semi-manually as they are very sensitive (changing the values of these coefficients has a significant impact on the total operating cost).
In order to make operations as efficient as possible, it was necessary to automate the modification process of optimizing the values of the coefficients. A modification module was thus developed for use in the Optimization Model. In this module, a customized search algorithm to optimize the values of coefficients was developed, inspired by the basic principles of Pattern Search.
The amount and number of times the step size should be varied depends on the amount of precision required. Once the step size has been changed, the optimizer is run again with the modified value of coefficients. If all possible directions have not been tried, the direction is modified and the optimizer is run again, and the entire process repeated until all the different step sizes have been tried for all directions without any improvement.
The modification process can take up to several days to run on a large data set, because it runs the optimizer hundreds of times. In order to make the modification process more efficient, it is first run with a smaller data set, to find the initial values of the coefficients. These coefficients are then used for the larger data set, to fine tune the initial values and further optimize the values of the coefficients. This makes it feasible to run the modification process even with large data sets within a day.
It is necessary to run this modification module only when there are changes in the company's operational policy. For instance, the company's policies on the importance of the wait time cost, home base and getting closer to home base factors, and fairness factors such as a tractor's maximum time away and maximum mileage might change. This change in the company's operational policy must be reflected in the functions where these factors are calculated and the modification process must then be re-run to optimize the values of the coefficients.
There are two main types of efficiency gains obtained from the Optimization Model. One of the gains is quantifiable, while the other is non-quantifiable. The quantifiable gains are obtained by comparison of the total operating cost obtained in the Optimization Model and a known reference model. The same input data set is used in both models, to determine the magnitude of savings that can be made in terms of the total operating cost. The operating cost is based on the four values of full miles, broker miles, empty miles, and wait time cost. Thirty days of operational data from a trucking company were used, and the average percentage changes in these values between their current Model and the Optimization Model were calculated. This percentage change in these four values represents the efficiency change in the operations when the Optimization Model is used.
The annual economic benefits of using the Optimization Model can be estimated by calculating the difference between the annual cost of existing operations and the annual cost determined by using the Optimization Model. The total annual cost of linehaul operations for a trucking company was calculated by using past operational data from January 2005 to December 2005. The percentage changes (that were previously calculated) were applied to the existing operational values to estimate the operational values for the same time period assuming the decisions had been made using the Optimization Model. The estimated operational values for the Optimization Model were used to calculate the total annual operating cost.
It is not possible to directly quantify the benefits of higher customer, planner, and owner-operator satisfaction obtained from better decision-making and scheduling. However, these factors can be estimated in time through a series of surveys and also by observing the trends in owner-operator, customer, and planner retention. Observing the trends may not be a very accurate way to estimate these benefits, due to the changing nature of the market conditions.
The Optimization Model is a hybrid model to be used as a decision support system, which was developed by integrating the exact and heuristic solution approaches. This decision support system is intended to make the majority of operational decisions, as well as to assist in making tactical and strategic decisions. The model primarily helps to make operational decisions by optimizing job assignments to tractors.
Tactical decisions such as determining the right number of owner-operator tractors at a terminal can be made by running the model with different scenarios (number of owner-operator tractors at a terminal) and choosing the scenario that yields minimum cost.
Strategic decisions include determining the location of a new terminal or changing the location of an existing terminal to a different one. It would be almost impossible to carry out such intensive analyses manually, but the model can justify either an existing terminal location change or adding a new terminal by running both scenarios with the same operational data and choosing the scenario that yields lower operating costs. If the current situation has a lower operating cost, then no changes should be made.
The Optimization Model uses fixed and variable input data. The fixed input data includes the distance matrix, list of terminal codes, and information on switch points between O-D terminal pairs. The variable input data (as discussed above) used in the Model is the data on available loads and tractors, which is the same information that planners and dispatchers use to make operational decisions. The output from the model is primarily the operational decision plan, but also includes statistical information on the various parameters used to make these decisions. The optimization model will be discussed further under the following headings:
Fixed input data is data that does not have to be updated on a daily basis, but changes only in such situations as when a terminal or switch point is added (or removed).
Each cell of the distance matrix contains the distance (in miles) between the O-D pair represented by that cell. A sample distance matrix is given in Table 1:
Given the nature of linehaul operations, the time taken to travel between an O-D pair can be estimated by dividing the distance between the two terminals by the average speed (as a fairly constant speed can be maintained while traveling between terminals). However in cases where this estimation is not precise enough for the application, an additional time matrix can be constructed to store the time it takes to travel between O-D pairs.
The switch points for a given O-D pair are determined using two files—one that contains the actual information on the switch points, and the other that contains a matrix where each cell stores line numbers to reference the correct O-D pair in the first file. The default order for a given pair of terminals is to have the origin first, and then the destination. However a terminal that occurs as the destination in one case may be the origin in another. The system of relative referencing using the line number matrix makes it possible to determine the correct switch points for any O-D pair, regardless of the order in which the terminals occur. A sample 10×10 line number matrix for the first 10 terminals is given in Table 2:
A “−1” at a cell implies that no switch points are possible in this case, because the origin and destination are the same terminal.
Consider the O-D pair 0-7, which has a “6” at that cell. The corresponding switch point(s) for that pair can be found by going to the corresponding line number (row 6) of Table 3 which contains sample switch point information for 22 O-D pairs:
The switch point information for the O-D pair 7-0 is the same as for 0-7, and is thus available at row 6 as well. The table of switch points contains information on the O-D pair, the number of switch points for that O-D pair, and the actual switch points. “−1” is used when an O-D pair that does not have any switch points. A load may have as many as 9 legs in this Model, and can be changed as per the requirements. The line number matrix, as well as the switch points table has to be suitably changed when a terminal or switch point is added.
A terminal may have one or more codes in addition to the standard 3-letter code assigned to the terminal. An additional fixed input file is used to match such code(s) to the standard code. For instance consider the Hartland terminal, which has the standard code HAR. A US fleet tractor based in Hartland may have the code UHA to indicate its home base, and a pre-scheduled tractor may have the code SHA.
For a given situation (of jobs and tractors), a cost matrix is created, where each cell of the matrix represents the cost of assigning the corresponding job to the corresponding tractor. An infeasible assignment is indicated by a large number (10000). Infeasible assignments as discussed above occur in the following cases:
For a problem of this size and complexity, the efficiency of the solution algorithm is crucial, to be able to obtain the solution in an acceptable length of time. In the iterative approach, feasible assignments are made for the given situation and the completion times (new relative available time for the next iteration) are calculated. The iterative approach also reduces the problem size for the next iteration by considering only incomplete (if at least one leg remaining to be assigned) and pre-assigned broker loads. The original MIP problem was reformulated as an assignment problem so that the Hungarian Assignment algorithm or equivalent could be utilized. The Hungarian Assignment algorithm or equivalent is one of the most efficient exact solution algorithms and can determine the job assignments to tractors within a reasonable period of time even for problems of extremely large size.
As shown in the mathematical model below, the assignment problem is subject to 3 constraints:
Cost calculation plays a crucial role in making efficient use of the assignment algorithm, which will be discussed in detail in the next sub-section. The mathematical formulation of the model is presented in equations (1) to (12) below:
Equation (1) is the objective function, which is minimizing the total cost of tractor assignments to jobs. Equations (2) and (3) ensure that every tractor and job is assigned only once. Equation (4) is the binary decision variable, which can take a value of 1 if an assignment is made and 0 otherwise. Equation (5) shows the cost calculation for the assignment of owner-operator tractors and broker tractors to jobs and dummy jobs. Equation (6) represents the calculation of Tij, which is the earliest time that tractor i can start job j. Equations (7) and (8) denote the calculation of wait time and wait cost of tractor i fo job j, respectively. Equation (9) is the constraint that, if the origin or destination of a load is in the US, the tractor assigned to the load has to be eligible to drive in the US. Equations (10) and (11) are the maximum mileage and maximum time away constraints for tractor i, respectively, which ensure that the tractor does not exceed these predetermined limits. Equation (12) is the constraint for job j, which ensures that the job is delivered before its deadline.
The cost matrix contains the costs of assigning a load to a company tractor and also to a broker tractor. As discussed previously, infeasible or unprofitable assignments are indicated by a big number such as 10000 at the corresponding cell(s) of the cost matrix. As can be seen in the mathematical model, the cost (Cij) for feasible assignments is calculated as:
The cost depends on the following three main factors:
The cost per empty mile is lower than the cost per full mile, and the Empty Mile Ratio (EMR) is used to reflect this in the cost calculation. EMR is the ratio between the cost per empty mile and the cost per full mile, and is set at 0.9 for one embodiment of this present invention after discussion with the executives of a trucking company. Full mile costs are also different between Canada and the US as the company pays different owner-operator rates for a US and Canadian mile. Wait costs are incurred when a tractor is waiting at a non-home based terminal for a job longer than a predetermined number of hours (a tractor is available but there is no available load). When such a wait cost occurs, it is added to the full and empty mile costs in order to calculate the actual operating cost.
The general structure of the cost matrix used in the Model is shown in Table 4:
It is assumed that there are 0 to l−1 available jobs (along the columns) and 0 to t−1 tractors potentially available to do these jobs (along the rows). To calculate the trade-off between using broker tractors and owner-operator tractors automatically, l rows are added after the owner-operator tractor rows to indicate broker tractors. The cost for a job to be done by any of the broker tractors is the same and is denoted by BCj. In order to be able to use the cost matrix in the Hungarian Assignment algorithm or equivalent, it has to be a square matrix so that every job and tractor can have exactly one assignment. Thus t dummy jobs (columns) are added after the available job columns. The cost for all these jobs is set as 0. There are typically between 400 and 650 available tractors and 500 and 800 available loads for this linehaul operation. Thus a typical cost matrix for this Model is over 1000×1000 in size.
In order to be able to obtain the solution in a reasonable length of time, it was crucial to reduce the problem size as much as possible for every successive iteration. Thus in every iteration, the original cost matrix is reduced to a smaller matrix (the EM) to use in the Hungarian Assignment algorithm or equivalent. For each of the jobs from 0 to l−1, the jobs are renumbered from 0 to ll-1 in each iteration where ll is the number of available loads for that iteration.
The loads that are eliminated from the original matrix for an iteration are those jobs (all the legs of the job) that have already been assigned to an owner-operator tractor and those jobs for which the broker tractor cost is lower than any of the owner-operator tractor costs. In the first case (if all the legs have been assigned), the job is considered complete and therefore there is no need to do any further processing for that job. In this present invention, the second case is referred to as a pre-broker assignment in this thesis and since such jobs cannot be done at lower costs using company resources (owner-operator tractors), they will be assigned to broker tractors when using the Hungarian Assignment algorithm or equivalent. Hence these jobs are also eliminated from the original matrix.
Consider a situation of 400 available tractors and 600 available jobs. The original cost matrix to be used in the Hungarian Assignment algorithm or equivalent would be a 1000×1000 matrix. The Hungarian Assignment algorithm or equivalent would then have to process one million cells. However using this iterative approach, the number of available loads in the last iteration may be lower than 100. Even if the number of loads is 100 in the last iteration, this reduces the size of the matrix to 500×500, which effectively reduces the number of cells that have to be processed by a factor of 4. Table 5 shows the structure of the reduced cost matrix which is used in the Hungarian Assignment algorithm or equivalent:
Reducing the size of the cost matrix in this manner leads to significant savings in terms of the computational time required to obtain the solution. Using the original cost matrix for processing in every iteration would increase the computational time by a factor of at least 10. The original load numbers of the jobs in the reduced matrix are kept track of by using an indexing array.
The EM is sent to the Hungarian Assignment algorithm or equivalent, which processes the matrix to produce the final reduced EM. The Decision Matrix (final assignments) is generally calculated based on this reduced EM to find a random (any) optimum solution, but in this Model, further processing is done on the reduced EM using heuristics to obtain the best optimum solution as will be discussed in the following sections.
The general steps of the Hungarian Assignment algorithm or equivalent are given by Dyalog Limited (2006) as follows:
It should be noted that the Model uses the final reduced EM (obtained from step 3 when an optimum assignment can be made from the reduced matrix) for further processing to find the best optimum solution. Each of the 6 main steps of the algorithm was coded as a sub-program (function).
Two major heuristic algorithms were developed to find the best optimum solution. The first one is to find all possible optimum solution combinations by grouping jobs and tractors in a certain manner, which will be explained in greater detail further below. The second heuristic algorithm determines the best (preferred) optimum solution in such a way that the company tractors (including owner-operators) end up at the preferred location(s), which will lower the operating cost not only for the next iteration but also for the next day.
The heuristic which groups all the optimum solution combinations can be summarized in the following steps and acts:
For practical reasons, the possible groups are determined only for the company tractors and real jobs. This is because it is clear that any remaining tractors that cannot be assigned within a group will be assigned to dummy jobs, which means that they will not be assigned any other jobs for the current iteration. Similarly any remaining job(s) that cannot be assigned within a group will be assigned to broker tractor(s), which means that those job(s) will not be assigned in the current iteration. Broker tractors are still required in the Hungarian Assignment algorithm or equivalent to calculate the trade-offs between using the company tractors and the broker tractors. Dummy jobs are also required because the matrix has to be square to make efficient use of the Hungarian Assignment algorithm or equivalent.
Uncovered zeroes for each row and column are counted again. The least number of zeroes is 2 for row numbers 2 and 5. Based on this, the second group (group number=1) contains row numbers 2 and 5 and corresponding column numbers 0 and 2. An important point to be observed is that when column numbers 0 and 2 are covered, the cells (1, 0) and (1, 2) are also covered. This indicates that although these cells contain a zero, an assignment cannot be made for these zeroes. Otherwise it would not be possible to ensure having an optimum solution.
As shown in
In a similar manner uncovered zeroes are counted for the third group (group number=2). Columns 1, 4, and 5 and rows 1, 4, and 6 are grouped and covered as shown in
In this simple example, the number of rows and columns in a group happened to be equal. However this is not necessarily the case for real data. If this process is done only for the available loads and tractors (excluding dummy jobs and broker tractors) and if the number of rows or columns is not equal, it would mean that the excess tractor(s) or job(s) would be assigned to dummy jobs or broker tractors respectively. For example if there are 3 jobs and 2 tractors in a group, there would be 6 possible optimum assignment combinations and for each combination one job would have to be assigned to a broker tractor.
Even for this very simple example the number of possible optimum solution combinations is 2×2×3!=24. For real data, the number of different possible optimum assignment combinations is between 1025 and 10100. These are only the different assignment combinations that have the same minimum cost. The total possible solution combinations can be approximated to be between 101000 and 10200000. In absence of the present disclosure, there is no method at present that finds all possible optimum combinations systematically in order to make trade offs between optimum combination efficiently using the Hungarian Assignment algorithm or equivalent because for most applications any optimum solution is sufficient. However, because this Model uses an iterative procedure where the outcome of one iteration affects the next iteration significantly, this heuristic was developed, in accordance with one embodiment of the present invention.
Once the optimum groups are determined, all the cells within each of these groups are weighted using functions, based on several criteria. Only the cells corresponding to the company tractors and real jobs are weighted. The final assignments are made based on the weighted factors. These final assignments represent the best optimum solution, and the decision (assignment) matrix is created based on this preferred solution.
The Effective Decision Matrix (EDM) is created with the same dimensions as the EM to store the weighted factors used in making the final assignments. The following factors and ratios are used to weight the cells:
Certain factors and/or ratios that a tractor qualifies for are multiplied by a base number (CHF) to determine the weighting of that cell. Specific formulae were developed to determine each of these factors and ratios, and these formulae will be discussed later in this subsection.
The steps and acts of the heuristic are as follows:
The calculation of the five factors introduced above will be discussed in the rest of this subsection. The formula used to calculate the WTR is:
Wait time is calculated as the time that a tractor waits for the corresponding load or half of the time that a load waits for the corresponding tractor. This is because a tractor's wait time is paid, whereas a load's wait time is not as important especially because if the load is an optimum group, it will be delivered before its deadline. WTRF (Wait Time Ratio Factor) is a predetermined coefficient having a value of 0.6, which reflects the importance of wait time relative to other factors.
HBF and CHF are predetermined coefficients, which have values of 0.76 and 100, respectively. HBF indicates the relative importance of a tractor going home amongst other factors. All weightings are calculated based on CHF, which is a predetermined coefficient that has a constant value. Since the weighting matrix is an integer matrix, the value of CHF can be simply increased, if the precision of the weighting is to be improved.
If the destination of one of the remaining legs of the load is the same as the home base of the corresponding tractor, then the waiting is done based on WTR and LHBF with the following formula:
LHBF is also a predetermined coefficient having a value of 0.5, which indicates the relative importance of the destination of a leg of a load being the same as the home base of the corresponding tractor.
If a tractor is at its home base, then GCR is 1. If its at another terminal other than its home base then GCR is calculated depending on whether the tractor is getting closer to or away from the home base by doing the corresponding job. If the destination of the leg of a job is closer to the tractor's home base than the current location of the tractor, then GCR is calculated as:
Where dis_cur_hb is the distance between the current location of the tractor and the tractor's home base, and dis_leg_dest_hb is the distance between the destination of the leg of the load and the tractor's home base. If the current location of the tractor is closer to its home base than the destination of the leg of the load, then the GCR is calculated as:
The value of GCR thus obtained is used to calculate the final value of GCR by using the formula GCR=1+(GCRF*GCR). GCRF is a predetermined coefficient having a value of 0.4, which reflects the relative importance of tractor getting closer to its home base. Higher GCR value indicates that the tractor is getting closer to its home base.
MR is calculated as MR=(1+MRF)−(MRF*max(MMR,MTAR)). It can be seen that the value of MR is based on MRF and the maximum of the Maximum Mileage Ratio (MMR) and the Maximum Time Away Ratio (MTAR). MRF is a predetermined coefficient having a value of 0.01, which indicates the relative importance of the Maximum Ratio amongst other factors. The value of MMR is calculated as:
Similarly the value of MTAR is calculated as:
MR would be 1 if the preferred maximum time away and maximum mileage limits are exceeded. The value of MR might be between 1 and 2 depending on MRF and the maximum of MMR and MTAR.
If none of the remaining legs of the load have a destination the same as the tractor's home base, then the weighting for that cell is calculated using the following formula:
The method used to establish the predetermined coefficients such as HBF, LHBF, GCRF, MRF, and WTRF will be discussed in greater detail under the Modification below.
Once the input data on the available tractors and loads is validated, the cost matrix is created for the jobs, company tractors, and also for the broker tractors and dummy jobs. This cost matrix is then used in the Hungarian Assignment algorithm or equivalent to obtain the final reduced EM. The heuristics developed are used to determine the different possible optimum solution groups and consequently the best optimum solution, and the final decision (assignment) matrix. The assignment matrix as discussed earlier contains the assignment of jobs to company tractors as well as to broker tractors.
All assignments are made based on this decision matrix, and the related fields for the jobs and tractors are updated. The fields that are updated for the available tractors that are assigned are the ETA, work time, current time away, running mileage, wait time, cost, current location, load number, and leg number. The fields that are updated for the available jobs that are assigned are the EPT, ETA, tractor number, and the number of legs that have been assigned.
Once these fields have been updated, all loads that have been assigned to broker tractors or that have unassigned legs remaining, as well as all the tractors are made available for the next iteration. Depending on the new list of available loads and tractors, the cost calculation and assignments are made for that iteration, and this process is repeated until no more assignments can be made. Obtaining the final output from the optimization model requires one complete run of the program, which consists of approximately 25 (±10) iterations.
An understanding of the key concepts of relative time and time step is required to appreciate the iterative procedure used. Thus an outline of the iterative procedure is given under the heading termination of the iterative procedure, after introducing these concepts under the following heading.
The jobs and tractors are available at different times, some of them available several hours from the current time and some available several hours previous to the current time. To address this issue for a given situation of jobs and tractors, the time of availability of the jobs and tractors is considered in terms of relative time, which is the time relative to the current time. It is calculated as relative time=(current date−date read in from available load/tractor information file)*24−(current time−time read in from available load/tractor information file). In this formula, the current time and the time read in are not in the usual time format, but are in decimal format. For example, a current time of 2:40 would be 2.66 (2+40/60) and an available time of 4:20 would be 4.33, obtained by converting the minutes to hours. Thus if the current and available dates are the same, the relative time would be 4.33-2.66=1.67, implying that the load/tractor will be available 1.67 hours from the current time. Similarly a tractor/load available before the current time would have a negative value for relative time. The dates read in are in Julian format, so the difference in dates multiplied by 24 gives the difference in hours in the formula. A relative time of 50.5 implies that the job/tractor will be available two days and two-and-a-half hours from the current time.
This simple concept of relative time alone was not sufficient to address certain other cases. Consider a tractor that has a relative available time of 5 (hours), and a job that has a relative available time of 20. It may be beneficial for the current iteration to assign the tractor to the job even if wait time has to be paid. However it is possible that in the following iterations there will be a load that has a relative available time of 5, and there will be other tractors available to do the job that has a relative time of 20. It is obvious that, in this case, it is not appropriate to make the assignment in that current iteration. The concept of time step was introduced to address this issue. The assignment of a job to a tractor is considered infeasible if the difference in their relative times is greater than the time step. Time step has a predetermined value, and after several trial runs was fixed to be 8 for this Model. The boundary for the time step is relative to the relative time. For example, the tractor having a relative time of 5 can only be considered to do jobs that do not have a relative time greater than 13. However if the relative time of another tractor is 15, the corresponding time step boundary would be 23.
Consider that the current leg of L2 to be processed has a relative time of 3.5, L2 is assigned to T1, L1 and T2 have not been assigned, and L2 has at least one more leg remaining. The relative time step for the next iteration for such a situation is shown in
It is to be noted that the concepts of relative time and time step were crucial to improving the operational efficiency and the role of these concepts in the iterative procedure is further detailed in the next subsection.
Once the optimization is done for the current situation (of jobs and tractors) at step 1410, the first condition that is checked at step 1420 is whether at least one company tractor has been assigned to a job. If at least one such assignment has been made, the optimization procedure of step 1410 is repeated for the next iteration.
At step 1430, the second condition is checked when not even one company tractor assignment has been made. The second condition checks whether there is at least one possible company tractor assignment, which was not made in the current iteration, because the tractor's availability was at least 8 hours (example value) before the job's availability. In this case the available relative time of all the tractors is increased by time step at step 1440, and it implies that no assignments can be made without performing this increment.
This iterative procedure is continued until neither of the conditions is satisfied. When the iterative procedure is terminated, the required output containing operational decisions are produced at step 1450 and the optimization program ends.
As discussed above, the modification module is only run when it is required to determine the fixed values of the coefficients for the factors and ratios used in the best optimum solution procedure. This is required only when there are changes in the company's operational policy on any of the preferred and/or fairness factors used. This change in the company's operational policy must be reflected in the functions, where these factors are calculated, and the modification module must be re-run to optimize the values of the coefficients.
The modification module does not include the process of determining the value of MR, due to the absence of the required data. This data includes the maximum time away and maximum mileage fields for each tractor, which will be determined by the company in the near future. The structure of the modification module has been framed such that the process of determining MR can easily be added to determine MR, when required. The company has to also associate a cost with this factor in the Optimization Model, so that the weighting for MR can be determined relative to other factors.
The general steps of the modification algorithm are as follows:
The step size changes according to the step change, and step change is set to 0 when there is an improvement. When there is no improvement in any of the directions for a given step change, then step change is incremented by 1. Step size is determined depending on whether the step change is odd or even.
The step size starts at 5% (meaning that the coefficient will be changed by that percentage in the current direction), is decremented using the formula step size=step size/(2stepchange) when the step change is even and incremented using the formula step size=step size*(2stepchange) when the step change is odd. The maximum value of step change in the module is set to 6 in the Model, but this limit can be increased if more precision is required, at the cost of increased computation time. The values that step size takes in order are: 5%, 2.5%, 10%, 1.25%, 20%, 0.625%, and 40%. When there is an 100 improvement, the step size is set to 5% again. Different step sizes had to be used because the coefficients do not have a linear relationship. This implies that while a step size of 10% may not lead to an improvement, a step size of 20% might.
This modification module is used to optimize the values of HBF, LHBF, GCRF, and WTRF. It determines the weighting for these factors relative to one another while making the final decisions. The structure of this module was developed such that any number of additional coefficients can be added and their values can also be optimized.
The Optimization program had to be restructured to accommodate the modification module, because the main program is treated as a function where the objective function (determining the total operating cost) for the modification module is evaluated. The modification process is done based on the results from the function (main program). Significant changes had to be made to the main program after inclusion of the modification module, because all the information on available loads, tractors, and their related fields and variables had to be reset to the original values every time a run was made.
It was a very cumbersome and time consuming task to optimize the values of these coefficients manually by trial-and-error. At least hundreds of runs would have had to be made and the coefficients modified manually after each run. It was also very difficult to modify the coefficients in a systematic manner manually, because they are inter-related. Even if the coefficients could have been optimized manually, a potential policy change would necessitate repeating this cumbersome task from the beginning every time such a change is made.
It is clear that the process of optimizing the coefficients had to be automated. Thus the modification module was developed to address this issue. The total operating cost is highly sensitive to even very small changes in these coefficients. It can be seen that this unique module has a significant contribution to the Optimization Model by optimizing the values of the coefficients that are used in it (the Optimization Model).
The Optimization Model has several supporting modules which are implemented as functions in the program. These functions are discussed briefly in this section, and are as follows:
There are also other supporting functions that are used in the program. However the processing done in these procedures is relatively simple and thus no flowcharts are provided for the following functions:
The sensitivity analysis shows the effect of the change in each of the four coefficients HBF, LHBF, GCRF, and WTRF on the total operating cost in the Optimization Model. As discussed earlier, these coefficients are used to weight the assignments to determine the best optimum solution. Since this is an iterative procedure, the outcome of one iteration has a significant impact on the iterations to follow, thus making these coefficients very sensitive to change.
It can be seen from the sensitivity analyses that the operating cost is minimal for the base values of the coefficients and increases with the corresponding percentage change in the coefficients.
It can be seen from Table 6 that 36.2% of available tractors are at their home base terminals on average before planning, while the rest are at non-home based terminals:
The average distance per tractor from its home base is 567 miles. It can also be seen that the average percentage of tractors that end up at their home base after planning using the Company Model is only 36.3% (as expected), whereas it is 42.1% after planning using the Optimization Model. The average distance per tractor from its home base terminal is reduced insignificantly to 563 miles (as expected) using the Company Model, while it is reduced to 453 miles (about 20% reduction relative to before planning) using the Optimization Model. This means that there is a 5.8% improvement in the number of tractors ending up at the home base and a −19.5% improvement (19.5% reduction) in the average miles away from the home base per tractor using the Optimization Model (relative to the Company Model).
As the results show, although only 5.8% more of the tractors end up at the home base using the Optimization Model, the average miles away from home base per tractor is reduced by 19.5%. This is due to the introduction of the Getting Close Ratio concept in the Optimization Model. It can be expected that there will be further improvement once the full implementation of the Model has been completed. This is because each subsequent run will start with a better placement of tractors, thus improving the final situation.
The non-quantifiable benefits of owner-operator and driver satisfaction would also increase when the above improvements are made because more tractors (and thus drivers) end up at their home base. Additional owner-operator and driver satisfaction is achieved by incorporating the fairness factors such as preferred maximum mileage and maximum time away into the Optimization Model.
The average violation of the LOS requirements in a trucking company is approximately between 1% and 5%. The two main contributors to this violation are operational mistakes due to human errors and miscalculations by the planners while planning. Some amount of operational mistakes cannot be avoided due to the human factor, but increased customer satisfaction can be achieved because planning errors are eliminated by using the Optimization Model (as all jobs are planned according to their deadlines).
A complexity analysis of the Optimization algorithms is presented in this section in order to show the increase in computation time when the sample size for available loads increases. The optimization is analyzed based on its three components, which are the cost calculation algorithm, the Hungarian Assignment algorithm or equivalent, and the assignment heuristics. The cost calculation algorithm is used to obtain the EM, which is used in the Hungarian Assignment algorithm or equivalent to obtain the final reduced matrix. The assignment heuristics include the heuristics used to determine the combinations of possible optimum solutions and also the heuristic used to determine the best optimum solution for final assignments.
The computation time for the optimization section of the program (excluding the input and output sections) is given in seconds in Table 7, for various sample sizes of available loads (the number of available tractors is assumed to be the same as the number of available loads):
It should be noted that the computation time is heavily dependent on the RAM available in the computer and also the processor speed. The timings presented in this section are from runs that were made on a computer that uses an Intel Xeon 3.06 GHz processor, with 2 gigabytes of RAM.
The complexity analysis for the three components showing the change in computation time in relation to an increase in the number of available loads, as well as for the entire optimization, is also presented in this section. The data points are represented by the “*” and the curve that fits this data is represented by a solid line (power line).
The operational data obtained from an example company for the year 2005 showed that their annual linehaul operational cost was approximately $133.88M (million). Analysis revealed that the maximum possible efficiency gain for the company linehaul operating cost in 2005 is 9.2% (±0.3%). This cannot be achieved in practice due to the nature of linehaul transportation operations, as the assumption to achieve maximum possible efficiency gain was having no empty miles, broker miles, and wait cost.
Quantitative validation of the company model was done by using the t-Test and this showed that the company model and manual planning are not statistically different from each other in terms of the operating cost as well as full miles, empty miles, broker miles, and wait cost. The ratios of each of the operating cost, full miles, empty miles, broker miles, and wait cost to the loaded miles were used in the t-Test for the statistical analysis since they describe the characteristics of the linehaul operations. These characteristics are the total operating cost for one loaded mile, the full mile for one loaded mile, the empty mile for one loaded mile, the broker mile for one loaded mile, and the wait cost for one loaded mile, respectively.
Efficiency improvements were quantified by using thirty datasets of available tractors and loads from September 2005 to March 2006 by comparing the company model results to the Optimization Model results. The efficiency gains in full miles, empty miles, broker miles, wait cost, and total operating cost were quantified as savings relative to the company model values. Savings is the difference between the company model values and Optimization Model values.
The average efficiency gains for full miles, empty miles, broker miles, wait cost, and operating cost were found to be −3.62%, 52.61%, 51.71%, 52.33%, and 4.82%, respectively. If the Optimization Model had been used in 2005, the corresponding projected lowest, average, and highest savings are approximately $6.18M, $6.45M, and $6.71M, respectively.
While the maximum possible efficiency gain is shown to be 9.2% assuming ideal conditions, the average efficiency gain with the Optimization Model is 4.82%. This suggests that the Optimization Model produces a result within 5% of the maximum possible global optimum. It is estimated that the Optimization Model produces a result within 2% to 3% of the real global optimum, since ideal conditions can never be achieved realistically.
Some operational improvements are anticipated which could improve the overall operational efficiency are addressed, in addition to linehaul planning and scheduling. The areas for potential improvements are identified as follows:
Dock managers currently evaluate the line haul trailer efficiency by predicting the volume of the trailer used (trailer utilization) before the trailer is sealed. This subjective method is prone to be inaccurate due to possible miscalculations or misleading benefits (operations managers might want a higher record of trailer utilization). The other method used to calculate the trailer utilization is to sum the weights of all the individual shipments. However, because the trailer generally cubes out before it weighs out, this method is also ineffective in predicting the trailer utilization.
In the current pricing policy the Equivalent Cubic Measurement is calculated assuming that 85% of the trailer can be filled. Since all dimensions and equivalent cube of individual shipments are known, it is possible to integrate this Equivalent Cubic Measure to evaluate trailer utilization and automate this process. At least 100% utilization should be achieved using this method and maybe more if there are shipments that are priced based on their weight. The main advantage of this method is its accuracy and independence from human subjectivity and that it can also be used as a verification tool for the pricing policy. It is recommended that transportation companies adopt this type of objective method to evaluate their trailer utilization.
Fork lifters, dock managers, and operations managers have certain objectives and priorities while executing their part of the trailer stripping and loading operation. It is clear that each party wishes to optimize their part of the operation, which is not necessarily the optimum for the whole operation. In order to optimize the overall operations, it is necessary to make operational decisions that need not necessarily be best for a single part of the operation. Determining a set of working procedures would significantly benefit the overall operational efficiency. The procedures can include how to make decisions with respect to cross-docking, configuration of gates in terminals, and mixing loads (weighed out and cubed out trailers).
Cross-docking would be very beneficial for short line haul and pickup and delivery operations in particular, but for long linehaul trailers cross-docking would increase the operating cost. Therefore it is recommended that for long linehaul trailers, the individual LTL shipments should be placed on the dock close to the trailer and the trailer should be loaded beginning with the pieces that would use as much volume as possible in the trailer. Optimization of the assignment of gates to trailers would reduce the conflicting forklift traffic on the dock and would also reduce the distance traveled by forklifts, which can improve the terminal's operational efficiency significantly.
Since there are significantly more trailers that cube out before they weigh out, there is a great opportunity to mix the weighed out trailers with the cubed out trailers to further utilize the shipping capacity. Even if there is another cubed out trailer going to the same destination at the smaller terminals mixing loads could be an option most of the time at larger terminals.
Some of the general operational suggestions will be outlined in the rest of this section. When there are two trailers with similar time windows, giving priority to the trailer with the most LTL shipments (since more customers are being served) would make the overall operation steadier and robust in the long run. However, when there is a trailer with a more sensitive time window, then the priority should be given to that trailer.
The trip is generally started once the tractors and loads are both available, which may result in increased wait time. This particular case can be prevented by starting the trip at an optimum time such that when tractors arrive at their destination there will be loads available for them to pick up, especially when their time windows permit this arrangement, which would result in decreased or no wait time.
The transportation companies would benefit from having a good IS to keep track of all required information for planning and operational analysis. It would also make it easier to integrate such an IS with the Optimization system, since data transfer for input from the IS and output from the Optimization system can be achieved easily. Using a common IS throughout all the carrier's operations would ensure compatibility between the various types of operations and increase their functionality, while being more cost effective by sharing resources.
The implementation of the Integrated Optimization System requires certain pre-implementation measures to be taken. These are discussed along with some general implementation suggestions in this section.
The parameters for contributors to the operating cost such as the cost per US mile, cost per Canadian mile, empty mile ratio, wait cost, and cost per broker mile should be modified in the Optimization Model by considering subsidized fuel price when the costing scheme for the company changes for any of the contributors.
As with any IS system, the quality of the output from the model is almost completely dependent on the input that is fed into it. The input data should have all the attributes required to reflect the actual operational situation. Although the data pre-processing and validation module makes sure that the number of errors in the input data is minimized, the core input data itself needs to be kept track of in a consistent manner. Such data includes the current time away, estimated time of availability and work time for the available tractors, the release times for loads along with their delivery times and the number of LTLs. It should also be ensured that at any given point in time, all available tractors must be present in the system and their status read as “available”. Once the data is recorded in such a manner and rigorous testing has been performed with this data, the next step would be to create the required user-interface to integrate the model into the existing setup at a company.
Having fair and preferred maximum mileage and maximum time away for the tractors would result in savings that cannot be quantified, but that could improve owner operator retention considerably.
It is not intended that the integrated optimized system be used as a stand-alone tool that will make all the operational decisions without human intervention. While the main strength of the system lies in its ability to perform calculations that are too complex to be done manually, it also necessitates educating the planners and dispatchers to interpret the output of the model in the right manner. A proper understanding of this output will enable them to make more efficient decisions based on it. For instance if a job is assigned to a broker by the Optimization system, the planners can intervene to check if this job can be done by a company shunt or P & D tractor instead.
The suggested final implementation of the Optimization system is discussed in the following section. The input to the Model is obtained directly from the IS (such as for example AS/400 for a particular company) and the processing is performed. The output is fed back to the IS as a pre-plan, which the planners can view on their dispatch screens. Once the assignments of jobs to tractors are confirmed by the planners, the respective tractor drivers are informed via the satellite communication systems provided.
Testing of the integrated system with the IS at a company to be done in parallel to manual planning, due to the nature of the transportation operations. Typically, P& D operations are done only during business days whereas linehaul operations are done on the weekends and holidays as well. Use of the integrated system to make daily operational decisions should commence only when all testing and interfaces for integration have been satisfactorily completed.
The following publications are included herein by reference:
The above-described embodiments of the present invention are intended to be examples only. Those of skill in the art may effect alterations, modifications and variations to the particular embodiments without departing from the scope of the invention, which is set forth in the claims.